Skip to main content
Governance only covers the AI you can see. Device telemetry extends Kelasa’s visibility to the machines your team works on: a lightweight agent on each enrolled device reports evidence that AI tools are present and in use — which apps, which MCP servers, how much Claude Code usage — so admins can see real adoption and spot shadow AI, without ever collecting anyone’s prompts, files, or browsing content. Everything lives under AI Usage in the sidebar (admin-only): a people roster, each person’s devices, and a per-device dashboard.

​
What admins see

Usage

Claude Code token and estimated-cost totals per device and per person, over 7- or 30-day windows, broken down by model — and by skill, agent, or plugin where that’s observable.

Inventory

Installed AI apps, configured MCP servers, and Claude Code skills, plugins, subagents, and hooks found on the device.

Browser AI visits

Visit counts to known AI domains from local browser history — counts per domain only, never URLs, titles, or search terms.

Alerts

Shadow-AI signals raised as observations with sourced evidence, which admins can acknowledge and resolve. Device online/offline status updates live.

​
Shadow-AI alerts

The backend raises three kinds of alerts. Each is an observation with evidence, never a verdict — the UI says “Personal / other-org account observed”, not “policy violation”:
  • Personal account — a device’s AI session is signed into a personal or other-org account, bypassing your organization’s governance and billing.
  • Unlisted MCP server — an MCP server is configured on the device that isn’t in your organization’s app list.
  • Agent silent — an enrolled device has stopped checking in, so its coverage can no longer be trusted.
Alerts carry timestamped evidence rows (metadata only) and move through open → acknowledged → resolved.

​
What is never collected

Telemetry is filtered on the device through a fail-closed allowlist before anything is stored or uploaded — fields not explicitly allowed are dropped. The agent never transmits:
  • Prompt text, responses, or message bodies
  • File contents, file paths, or source code
  • Full URLs, page titles, or page content
  • Tool inputs and results, command arguments, or error text
Identifiers that are needed for comparison (account emails, config fingerprints) are one-way hashed with a per-device key before upload — they can be matched, never read back. Browser history is checked only against a fixed allowlist of AI domains; nothing else in the history is read out.

​
Honest coverage

The dashboard distinguishes not observed from not used. If a collector can’t read something — for example, the agent hasn’t been granted Full Disk Access on macOS — the device page reports that coverage gap instead of showing a reassuring zero.

​
Availability

The agent runs as a background service and currently supports macOS; Windows and Linux are planned. See Enroll a device for setup.